[tmva][sofie] Restructure emitted code to be differentiable with Clad

The idea of this commit is to refactor the `doInfer()` function that
implements the inference from a member function of the `Session` struct
to a free function that takes the `Session` by `const`-reference.

Author: guitargeek

tmva/sofie/inc/TMVA/RModel.hxx

@@ -34,8 +34,14 @@ private:
    std::vector<std::string> fDimShapeNames; // parameter names used to define the shapes
    std::vector<std::string> fOutputTensorNames;
    std::vector<std::string> fInputTensorNames; // input tensor names using ONNX order
+   std::vector<std::string> fPointerMemberNames;
 
+   inline std::string AddTensorMember(std::string const &name) {
+       fPointerMemberNames.push_back(name);
+       return "tensor_" + name;
+   }
 
+   bool IsInputTensorShapeParam(std::string const &name) const;
 
    std::vector<std::unique_ptr<ROperator>> fOperators;
 

tmva/sofie/inc/TMVA/ROperator_Gemm.hxx

@@ -393,9 +393,12 @@ namespace SOFIE{
              << (fAttrTransB ? "true, " : "false, ")
              << (fAttrTransA ? "true, " : "false, ")
              << n << ", " << m << ", " << k << ", ";
-            out << std::setprecision(std::numeric_limits<float>::max_digits10) << fAttrAlpha << ", tensor_" << fNB;
+            // TODO: the cast to (float *) is not needed here from the C++ language perspective (the arguments to
+            // Gemm_Call are const already), but Clad bug https://github.com/vgvassilev/clad/issues/1721 is requiring
+            // us to do this cast to keep Clad working. Remove this hack once the Clad issue is fixed.
+            out << std::setprecision(std::numeric_limits<float>::max_digits10) << fAttrAlpha << ", (float*)tensor_" << fNB;
             if (extraB) out << " + " << opName << "_B_offset";
-            out << ", tensor_" << fNA;
+            out << ", (float*)tensor_" << fNA; // TODO: same here
             if (extraA) out << " + " << opName << "_A_offset";
             out << ", " << std::setprecision(std::numeric_limits<float>::max_digits10) << fAttrBeta << ",";
             // in the case of bias and no broadcasting needed

tmva/sofie/inc/TMVA/ROperator_NonZero.hxx

@@ -101,16 +101,16 @@ public:
          }
       }
    }
+
    std::string GenerateSessionMembersCode(std::string /*opName*/) override {
       if (fIsOutputConstant) return "";
       // define output value used as max non zero with max size = input shape * N
       auto inputLength = ConvertDimShapeToLength(fShapeX);
       std::stringstream out;
-      out << SP << "size_t v_NonZero_" << fNX << " = " << inputLength << ";\n";
+      out << SP << "size_t fV_NonZero_" << fNX << " = " << inputLength << ";\n";
       return out.str();
    }
 
-
    std::string Generate(std::string opName) override {
       if (fIsOutputConstant) {
          return "";
@@ -133,7 +133,7 @@ public:
 
       // loop on input indices
       out << SP << "size_t offset_" << opName << " = 0;\n";
-      out << SP << vnonzero << " = 0;\n";
+      out << SP << "size_t " << vnonzero << " = 0;\n";
       for (size_t j = 0; j < dims; j++) {
          std::string index = "i_" + std::to_string(j);
          for (size_t k = 0; k <= j; k++) out << SP;

tmva/sofie/inc/TMVA/SOFIE_common.hxx

@@ -681,16 +681,6 @@ void col2im(const Dtype* data_col, const int channels,
   //std::cout << "finishing col2imp" << std::endl;
 }
 
-// Used at the end of infer() to fill the return object.
-template <class T>
-void FillOutput(T const *arr, std::vector<T> &out, std::size_t n)
-{
-   out.resize(n);
-   for (std::size_t i = 0; i < n; ++i) {
-      out[i] = arr[i];
-   }
-}
-
 }  // end namespace UTILITY
 
 namespace BLAS{